Equipment fault diagnosis recommendation method based on knowledge graph and multidimensional correlation indicators

Abstract

The present invention provides an equipment fault diagnosis recommendation method based on a knowledge graph and multidimensional correlation indicators, comprising: converting fault text into triple-structured data based on an equipment fault knowledge graph ontology model to construct an equipment fault knowledge graph; constructing a multidimensional correlation indicator for fault modes to convert explicit feedback in historical fault diagnosis data into implicit feedback; obtaining fault modes that are strongly correlated with abnormal phenomena from historical fault diagnosis data based on the multidimensional correlation indicator as seed fault modes; using the seed fault mode as a starting point, iteratively propagating through the fault knowledge graph to obtain a ripple set, calculating the association probability between abnormal phenomena and candidate fault modes based on the ripple set, accumulating responses of each order based on the association probability, and outputting a predicted diagnosis probability. The present invention reveals the strength of the diagnostic relationship between abnormal phenomena and fault modes based on the multidimensional correlation indicator, and on this basis adopts a flexible reasoning mechanism to improve diagnostic accuracy and adaptability.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment fault diagnosis, and relates to an equipment fault diagnosis recommendation method based on a knowledge graph and multidimensional correlation indicators. Background Art

[0002] Equipment refers to mechanical devices and systems with specific functions and designed to perform specialized tasks. They are widely used in key sectors such as national defense, manufacturing, and transportation, fulfilling mission assurance, improving production efficiency, and performing specialized operations. During equipment operation, the effective diagnosis and timely resolution of faults are directly related to operational efficiency and the reliability of mission completion.

[0003] In recent years, knowledge graphs have attracted widespread attention for their information integration and reasoning capabilities in fault diagnosis, and many researchers have conducted in-depth exploration in this area. However, current diagnosis recommendation methods based on knowledge graphs still have some shortcomings when applied to complex equipment fault diagnosis:

[0004] First, the fault knowledge graphs constructed by existing research often have a relatively simple structure and mainly focus on machine factors, but lack consideration of the two key elements of people and the environment, making it difficult to fully reveal the root cause relationship of complex faults.

[0005] Second, traditional knowledge graph embedding and path-based reasoning methods have insufficient interpretability of recommendation results in practical applications and are heavily dependent on the limitations of predefined meta-paths, making it difficult to meet the flexibility and comprehensiveness requirements of equipment diagnosis.

[0006] Third, to address the problem of insufficient interpretability of recommendation results, the RippleNet recommendation algorithm based on knowledge graph embedding and path hybrid can be considered. However, when applied to fault diagnosis, the existing RippleNet recommendation algorithm is difficult to accurately reflect the correlation between different fault modes and abnormal phenomena, further affecting the diagnostic accuracy.

[0007] Therefore, how to provide an equipment fault diagnosis recommendation method that can further improve the diagnostic accuracy is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] In light of this, this paper proposes a method for recommending equipment fault diagnosis based on a knowledge graph and multidimensional correlation indicators. This method reveals the strength of the diagnostic relationship between anomalies and fault modes in historical diagnostic information. Based on this, a flexible reasoning mechanism is employed to iteratively propagate fault information within the knowledge graph, expanding and dynamically mining the complex relationships between emerging anomalies and potential fault modes layer by layer. This iterative propagation approach enables in-depth exploration of fault modes across a wider range of graph levels, thereby improving diagnostic accuracy and adaptability and meeting the practical needs of complex equipment fault diagnosis.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] The present invention discloses a method for recommending equipment fault diagnosis based on a knowledge graph and multidimensional correlation indicators, comprising the following steps:

[0011] S1: Based on the abnormal phenomena of equipment units and the fault logic of human factors, machine factors, and environmental factors, the semantic types and semantic relationships in the field of equipment failure are defined, and the equipment failure knowledge graph ontology model is constructed;

[0012] S2: Obtain fault text from equipment failure cases, and convert the fault text into triple-structured data using a knowledge extraction algorithm based on the semantic types and semantic relationships in the equipment failure knowledge graph ontology model to construct an equipment failure knowledge graph; the equipment failure cases include historical fault diagnosis data;

[0013] S3: Constructing a multi-dimensional correlation index between abnormal phenomena and fault modes, using the multi-dimensional correlation index to optimize the correlation representation between abnormal phenomena and fault modes, and converting explicit feedback in historical fault diagnosis data into implicit feedback;

[0014] S4: Receive the input abnormal phenomenon, and based on the correlation representation, obtain the fault mode whose correlation meets the preset conditions from the historical fault diagnosis data as the seed fault mode; construct a diagnosis recommendation model, take the seed fault mode as the starting point, perform iterative propagation in the fault knowledge graph to obtain a ripple set, obtain the fault mode in the propagation process based on the ripple set, and use the seed fault mode and the fault mode in the propagation process together as candidate fault modes; calculate the association probability between the input abnormal phenomenon and the candidate fault mode based on the ripple set, accumulate each order response based on the association probability, output the predicted diagnosis probability, and determine the diagnosis result.

[0015] Preferably, the equipment fault knowledge graph ontology model has a multi-level structure, including:

[0016] Unit layer, used to describe the spatial location and structural association information of equipment, systems and components;

[0017] The logic layer is used to describe the fault characteristics and mechanisms and analyze the fault causes, including any one or more combinations of the following: human factors, machine factors, and environmental factors;

[0018] Maintenance layer, used to define repair measures and solutions.

[0019] Preferably, the semantic types in the equipment fault knowledge graph ontology model include any combination of the following: equipment name, system name, component name, failure mode, abnormal phenomenon, human factor, machine factor or environmental factor; the semantic relationship includes any combination of the following: composition, belonging, located, caused, induced, represented or taken.

[0020] Preferably, the equipment failure case of S2 includes structured text information and unstructured text information; wherein,

[0021] Structured text information includes historical fault data;

[0022] Unstructured text information includes professional technical documents, troubleshooting records, and maintenance manuals.

[0023] Preferably, a pre-trained BERT-BiLSTM-CRF model is used to perform named entity recognition on the unstructured text information; and a pre-trained Chinese-BERT-wwm-ext model is used to extract the relationship between fault-related entities from the unstructured text information.

[0024] Preferably, the S3 includes: obtaining abnormal phenomena and corresponding confirmed fault modes from historical fault diagnosis data, and using the fault frequency, fault level and impact range corresponding to the fault mode as multi-dimensional correlation indicators; wherein,

[0025] The fault frequency is used to record the number of times the abnormal phenomenon is diagnosed as a corresponding fault mode;

[0026] The fault level is used to assess the severity of the fault according to the impact of the fault mode;

[0027] The impact range is used to measure the impact of the failure mode on various components in the system.

[0028] Preferably, the step of optimizing the correlation representation between abnormal phenomena and failure modes using the multi-dimensional correlation index in S3 includes:

[0029] Standardizing the multi-dimensional correlation indicators to the same dimension;

[0030] Substitute the standardized multi-dimensional correlation index into the Failure Mode Association Evaluation Index (FMAEI) calculation formula to measure the correlation strength between abnormal phenomena and failure modes. The calculation formula is as follows:

[0031] ;

[0032] Where α1, α2, and α3 are the weight coefficients of the three dimensions of fault frequency, fault level, and impact range, and the sum of the weight coefficients is 1; S f , Ss, S i They are respectively the fault frequency, fault level and impact range data after standardized processing.

[0033] Preferably, in S4, a diagnosis recommendation model is constructed based on the RippleNet recommendation algorithm.

[0034] Preferably, the S4 comprises the following steps:

[0035] S41: calculating the correlation representation between the abnormal phenomenon and the fault mode according to the historical fault diagnosis data, and obtaining the fault mode that meets the preset conditions in the history of the input abnormal phenomenon as a seed fault mode;

[0036] S42: placing the seed fault pattern into the equipment fault knowledge graph as a starting point, obtaining a related ripple set through iterative propagation, obtaining a fault pattern during the propagation process based on the ripple set, and using the seed fault pattern and the fault pattern during the propagation process as candidate fault patterns;

[0037] S43: Calculating the association probability between the input abnormal phenomenon and the candidate fault mode based on the head entity and relationship of the ripple set;

[0038] S44: accumulating each order response of the ripple set layer by layer according to the associated probability, accumulating each order response, combining the corresponding candidate fault mode, outputting a predicted diagnosis probability, and determining a diagnosis result according to the predicted diagnosis probability.

[0039] Preferably, the step of calculating the predicted diagnosis probability between the abnormal phenomenon and the candidate failure mode in S4 includes:

[0040] Obtain the 1-hop ripple set of the candidate fault mode v and the input abnormal phenomenon w in the equipment fault knowledge graph , by combining the candidate failure mode v with the head entity and the relationship in the corresponding triple For comparison, Each triple in Assign an association probability :

[0041] ;

[0042] Where, , For the relationship r i and head entity h i Embedded, is the embedding of candidate fault mode v, d is the dimension of embedding, p i Indicates v and h i The degree of correlation between The number of triplets in , softmax() is the softmax function;

[0043] Pick All the tail entities in the and the corresponding association probability p i Perform weighted summation to obtain the diagnostic history V of the abnormal phenomenon w w The first-order response vector to the candidate failure mode v :

[0044] ;

[0045] Where, For the tail entity t i Embedding;

[0046] For the first-order response vector In the ripples of the anomaly w Perform iterative propagation to obtain the high-order response of the abnormal phenomenon w , the response of the abnormal phenomenon w to the candidate fault mode v is expressed as the sum of the responses of each order:

[0047] ;

[0048] Where, is the H-th order response of the abnormal phenomenon w, where H is a positive integer greater than 2;

[0049] Combining the abnormal phenomenon w and the candidate failure mode v, we can get the predicted probability:

[0050] ;

[0051] Where, is the sigmoid function, is the embedding of the abnormal phenomenon w;

[0052] All candidate failure modes are screened according to the predicted probabilities to obtain a recommended list of confirmed results.

[0053] As can be seen from the above technical solution, compared with the existing method of relying on predefined reasoning paths to achieve diagnosis through simple search and matching, the beneficial effects of the present invention include:

[0054] (1) In the process of constructing the equipment fault knowledge graph, the present invention effectively enhances the comprehensiveness and accuracy of the graph by introducing multi-dimensional information of the three elements of people, machines, and environment, making fault diagnosis more suitable for actual application scenarios and making up for the deficiency of traditional knowledge graphs that focus too much on machine factors.

[0055] (2) The present invention provides a more detailed and comprehensive evaluation of the relationship between fault modes and abnormal phenomena through multi-dimensional correlation indicators, dynamically revealing the strength of the correlation between abnormal phenomena and fault modes in historical diagnostic information, and avoiding the limitation of insufficient correlation identification in traditional methods.

[0056] (3) The present invention automatically associates the input abnormal phenomena with strongly correlated fault modes through the diagnosis recommendation model, and conducts in-depth analysis based on the multi-level association relationships in the knowledge graph. This not only effectively reduces the burden of troubleshooting on maintenance personnel, but also improves the diagnosis speed and accuracy through a flexible reasoning mechanism, greatly enhancing the efficiency and applicability of equipment fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only embodiments of the present invention. Those skilled in the art can also derive other drawings based on the provided drawings without inventive effort.

[0058] Figure 1 A flowchart of a method for recommending equipment fault diagnosis based on a knowledge graph and multi-dimensional correlation indicators provided by an embodiment of the present invention;

[0059] Figure 2 A schematic diagram of the equipment fault knowledge graph ontology model structure provided by an embodiment of the present invention;

[0060] Figure 3 A schematic diagram of historical diagnostic information for the "insufficient power output" abnormal phenomenon provided by an embodiment of the present invention;

[0061] Figure 4 Schematic diagram of the iterative propagation of the "injector 1 failure" fault mode in the fault knowledge graph provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0063] like Figure 1 As shown, an embodiment of the present invention provides an equipment fault diagnosis recommendation method based on a knowledge graph and multi-dimensional correlation indicators. First, a multi-level equipment fault knowledge graph is constructed that comprehensively considers the three elements of human, machine and environment. Second, a fault mode correlation evaluation index containing multi-dimensional fault characteristics is used to transform explicit feedback in historical diagnostic information. Finally, a diagnosis recommendation model is constructed in combination with the RippleNet recommendation algorithm with efficient association reasoning capabilities, thereby achieving accurate diagnosis of equipment faults.

[0064] The steps include:

[0065] S1: Based on the abnormal phenomena of equipment units and the fault logic of human factors, machine factors, and environmental factors, the semantic types and semantic relationships in the field of equipment failure are defined, and the equipment failure knowledge graph ontology model is constructed;

[0066] S2: Obtain fault text from equipment failure cases. Based on the semantic types and semantic relationships in the equipment failure knowledge graph ontology model, the fault text is converted into triple structured data using a knowledge extraction algorithm to construct an equipment failure knowledge graph. Equipment failure cases include historical fault diagnosis data.

[0067] S3: Construct a multi-dimensional correlation index between abnormal phenomena and fault modes, use the multi-dimensional correlation index to optimize the correlation representation between abnormal phenomena and fault modes, and convert the explicit feedback in historical fault diagnosis data into implicit feedback;

[0068] S4: Receive the input abnormal phenomenon, and based on the correlation representation, obtain the fault mode whose correlation meets the preset conditions from the historical fault diagnosis data as the seed fault mode; build a diagnosis recommendation model, take the seed fault mode as the starting point, iteratively propagate in the fault knowledge graph to obtain the ripple set, obtain the fault mode in the propagation process based on the ripple set, and use the seed fault mode and the fault mode in the propagation process as candidate fault modes; calculate the association probability between the input abnormal phenomenon and the candidate fault mode based on the ripple set, accumulate the responses of each order based on the association probability, output the predicted diagnosis probability, and determine the diagnosis result.

[0069] In one embodiment, equipment fault diagnosis involves a large amount of complex and diverse knowledge, such as the equipment's structure, fault modes, and their causes. Organizing this knowledge in a computable and reasonable manner requires precise ontology construction. In order to systematically and comprehensively describe equipment faults and related information, we start from three key questions regarding the fault event: "Where is this fault located?", "Why did this fault occur?", and "What maintenance measures should be taken for this fault?", and construct an ontology structure with three levels: unit, logic, and maintenance. Among them:

[0070] Unit layer, used to describe the spatial location and structural association information of equipment, systems and components;

[0071] The logic layer is used to describe the fault characteristics and mechanisms and analyze the fault causes, including any one or more combinations of the following: human factors, machine factors, and environmental factors;

[0072] Maintenance layer, used to define repair measures and solutions.

[0073] In this embodiment, considering that equipment failure is usually the result of the interaction of multiple factors, in the analysis of the root cause of the failure, based on the perspective of "man-machine-environment" system engineering, the three major factors of human factors, machine factors, and environmental factors are unified and modeled in the equipment failure knowledge graph, comprehensively revealing the cause of the failure and providing a scientific basis and means for equipment fault diagnosis and maintenance. The overall equipment failure knowledge level division is shown in Table 1, and the relationship between each semantic type is as follows: Figure 2 shown.

[0074] Table 1 Classification of equipment failure knowledge levels

[0075]

[0076] In one embodiment, the semantic types in the equipment failure knowledge graph ontology model include any combination of the following: equipment name, system name, component name, failure mode, abnormal phenomenon, human factor, machine factor or environmental factor; the semantic relationship includes any combination of the following: composition, belonging, located, caused, induced, represented or taken.

[0077] In one embodiment, the equipment failure case of S2 includes structured text information and unstructured text information; wherein,

[0078] Structured text information includes historical failure data. This data is stored in a table format and contains information such as failure mode, faulty component, cause, and corresponding repair measures. Because this data has clear structured features, it can be directly converted into triples in the knowledge graph.

[0079] It should be noted that the historical fault diagnosis data in equipment failure cases also uses structured data, but it does not participate in the construction of the fault knowledge graph. Its role is to serve as the judgment basis for diagnosis recommendations together with the knowledge graph recommendation data.

[0080] Unstructured text information includes professional technical documents, troubleshooting logs, and maintenance manuals. Professional technical documents cover equipment operating instructions, technical standards, and other content, detailing the equipment's structure and operating principles. Troubleshooting logs are logs generated by field technicians during the diagnosis and troubleshooting process. Maintenance manuals provide operating specifications and solutions for equipment maintenance and repair. These data, such as professional technical documents, troubleshooting logs, and maintenance manuals, contain a large amount of unstructured text information, which needs to be converted into structured data through knowledge extraction technology to provide a foundation for the subsequent construction of knowledge graphs.

[0081] It is necessary to explain the historical fault diagnosis data and historical fault data:

[0082] Historical fault data refers to a structured knowledge base centered on fault modes, including information such as abnormal phenomena, fault modes, faulty components, the system to which they belong, fault levels, fault causes, and corresponding maintenance measures, as shown in Table 2.

[0083] Table 2 Example of historical fault dataset

[0084]

[0085] Historical fault diagnosis information refers to the information obtained from diagnosing a specific abnormal phenomenon. It includes the abnormal phenomenon, fault mode, fault frequency, fault severity, and impact range, and is stored in a five-column table. Fault frequency, fault severity, and impact range are explicit feedback, as shown in Table 3.

[0086] Table 3 Example of historical fault diagnosis dataset

[0087]

[0088] The training data of the diagnosis recommendation model includes: knowledge graph triple structure data and historical fault diagnosis data.

[0089] In one embodiment, a deep learning model is used to perform named entity recognition and relationship extraction on unstructured fault text, such as troubleshooting records and maintenance manuals, to obtain structured triplet data related to equipment faults. This embodiment uses a pre-trained BERT-BiLSTM-CRF model to perform named entity recognition on unstructured text information; and a pre-trained Chinese-BERT-wwm-ext model to extract relationships between fault-related entities in unstructured text information.

[0090] In this embodiment, the BERT-BiLSTM-CRF model is used for named entity recognition. The specific steps are as follows:

[0091] S201: Use annotation tools to annotate the fault text and use the BIO annotation method to mark the entity boundaries in the text. The BIO method can effectively mark whether each word is part of an entity and clearly identify the entity's start position (B), middle position (I), or non-entity position (O), forming a standardized annotation dataset.

[0092] S202: Input the labeled dataset into the BERT-BiLSTM-CRF model for training. The BERT layer generates contextual semantic representations of words, the BiLSTM layer further extracts sequence features, and the CRF layer optimizes entity boundaries to ensure the consistency and accuracy of entity recognition results.

[0093] S203: We used precision, recall, and F1 score as evaluation metrics, and improved the model's robustness and generalization capabilities by adjusting its hyperparameters. The F1 score is a metric used in information retrieval and statistics to measure the accuracy of binary classification models. It is defined as the harmonic mean of precision and recall.

[0094] S204: Apply the trained model to new fault text, automatically extract entity data from it, and organize it into a structured data table to provide data support for knowledge graph construction.

[0095] In this embodiment, the Chinese-BERT-wwm-ext model is used to extract the relationship between fault-related entities. The specific steps are as follows:

[0096] S211: Perform relationship annotation on entity pairs in the training data to form a relationship annotation dataset to provide training data for model learning.

[0097] S212: Use the Chinese-BERT-wwm-ext model for training, enhance the model's ability to understand semantic associations in text through full-word masking technology, and perform relationship classification learning on entity pairs in the annotated data.

[0098] S213: Apply the trained model to new fault text to automatically identify the association relationship between entities and output structured triples containing "entity 1-relationship-entity 2", providing high-quality data support for knowledge graph construction.

[0099] In one embodiment, to efficiently manage the constructed equipment failure knowledge graph, this embodiment uses the Neo4j graph database as the storage system. Specifically, the structured failure data obtained through knowledge extraction is saved in .CSV file format and then written to the neo4j graph database using Cypher, thereby constructing the equipment failure knowledge graph.

[0100] Neo4j graph database uses graph structures (nodes and edges) to store data. Compared with traditional relational databases, it can more intuitively and flexibly represent entities and their relationships. It is particularly suitable for representing complex multi-dimensional relationships and can greatly improve the efficiency of queries and reasoning.

[0101] Using the Cypher language, structured fault triples are written to the Neo4j graph database, forming a knowledge graph of equipment failures. Cypher's efficient and intuitive graph data manipulation capabilities support complex queries, facilitating in-depth searches for fault modes, causes, and related information. Leveraging the characteristics of graph databases, queries for fault-related information after diagnostic results are more efficient, ensuring rapid identification of the root cause of equipment failures and related information.

[0102] In one embodiment, S3 includes: obtaining abnormal phenomena and corresponding confirmed fault modes from historical fault diagnosis data, and using the fault frequency, fault level, and impact range corresponding to the fault mode as multi-dimensional correlation indicators; wherein,

[0103] Fault frequency is used to record the number of times an abnormal phenomenon is diagnosed as the corresponding fault mode;

[0104] Fault level is used to assess the severity of a fault based on the impact of the fault mode (such as downtime caused by the fault, difficulty of repair, etc.);

[0105] Impact range is used to measure the impact of a failure mode on various components in the system.

[0106] In one embodiment, the purpose of S3 is to calculate the Failure Mode Association Evaluation Index (FMAEI), which reveals the strength of the association between abnormal phenomena and failure modes, thereby providing accurate prior knowledge for the diagnosis recommendation model. The steps of optimizing the association between abnormal phenomena and failure modes using multi-dimensional association indicators include:

[0107] Standardized processing steps:

[0108] After obtaining historical fault diagnosis data, it will be standardized to ensure that the data of different dimensions have a unified scale and comparability. First, the mean μ and standard deviation σ of each dimension are calculated. The calculation formula is as follows:

[0109] ;

[0110] ;

[0111] Where x i is each data point in that dimension, and N is the total number of data points in that dimension.

[0112] Each data point is then normalized to obtain the normalized data S, so that the data in each dimension has zero mean and unit standard deviation. The normalization formula is as follows:

[0113] ;

[0114] Through standardization, the numerical values ​​of the three dimensions of fault frequency, fault level and impact range are ensured to be in the same dimension, avoiding the dominant influence of a certain dimension on the diagnostic results.

[0115] Steps for weighted calculation of failure mode relevance assessment index:

[0116] Substitute the standardized multi-dimensional correlation index into the Failure Mode Association Evaluation Index (FMAEI) calculation formula to measure the correlation strength between abnormal phenomena and failure modes. The calculation formula is as follows:

[0117] ;

[0118] Where α1, α2, and α3 are the weight coefficients of the three dimensions of fault frequency, fault level, and impact range, and the sum of the weight coefficients is 1; S f , Ss, S i They are respectively the fault frequency, fault level and impact range data after standardized processing.

[0119] It should be noted that this embodiment converts explicit feedback into implicit feedback, using multi-dimensional comprehensive indicators to process the explicit feedback in historical fault diagnosis information. In this way, explicit feedback information is converted into implicit feedback through multi-dimensional comprehensive indicators, thereby revealing the strength of the correlation between abnormal phenomena and fault modes.

[0120] In one embodiment, S4 constructs a diagnosis recommendation model based on the RippleNet recommendation algorithm. The algorithm captures the complex correlation information in the equipment fault knowledge graph through a multi-order response propagation mechanism, and ultimately calculates the diagnosis probability between abnormal phenomena and fault modes.

[0121] In one embodiment, S4 includes the following steps:

[0122] S41: Calculate the correlation between abnormal phenomena and fault modes based on historical fault diagnosis data, and obtain the fault modes that meet the preset conditions in the history of the input abnormal phenomena as seed fault modes; when this step is specifically executed, the fault modes with strong correlation are regarded as seed fault modes.

[0123] S42: After inputting the abnormal phenomenon data, the fault mode with strong correlation in the historical diagnostic information will be used as the starting point of the model. Through the multi-level association relationship in the fault knowledge graph, the model obtains the entities related to the abnormal phenomenon layer by layer through an iterative propagation mechanism to form a ripple set. Based on the ripple set, the fault mode in the propagation process is obtained, and the seed fault mode and the fault mode in the propagation process are used together as candidate fault modes to participate in the subsequent association probability calculation. When this step is specifically executed, starting from the abnormal phenomena that have appeared in the equipment history and the seed fault mode determined by the corresponding diagnosis, the relevant entities and ripple sets are obtained by propagating layer by layer through the links in the equipment fault knowledge graph.

[0124] S43: Calculate the association probability between the input abnormal phenomenon and the candidate fault mode based on the head entity and relationship of the ripple set.

[0125] S44: Accumulate the order responses of the ripple set layer by layer based on the association probability and the tail entity of the ripple set. This step simulates the process of fault propagation based on the connection information of the nodes in the graph, thereby obtaining possible fault modes and related fault causes. The model accumulates the responses of each order, combines the corresponding candidate fault modes, and outputs the predicted diagnosis probability between the abnormal phenomenon and the fault mode. The probability reflects the diagnostic confidence that a certain abnormal phenomenon is diagnosed as a certain fault mode. When this step is specifically executed, a prediction function is constructed to determine whether there is a potential diagnostic relationship between the abnormal phenomenon and the fault mode. The model form is:

[0126] ;

[0127] Where, represents the probability that the abnormal phenomenon w is diagnosed as the failure mode v, Represents the model parameters of the function F.

[0128] In one embodiment, the step of calculating the predicted confirmed probability between the abnormal phenomenon and the candidate failure mode in S4 includes:

[0129] S410: Obtain the one-hop ripple set of the candidate fault mode v and the input abnormal phenomenon w in the equipment fault knowledge graph , by combining the candidate failure mode v with the head entity h i and the relation r in the corresponding triple i For comparison, Each triplet (h i ,r i,t i ) assigns an association probability p i :

[0130] ;

[0131] Where, , For the relationship r i and head entity h i Embedded, is the embedding of candidate failure mode v, express The transposed embedding matrix, d is the embedding dimension, p i Indicates v and h i The degree of correlation between Softmax is a normalization function that limits the calculation results to between 0 and 1 and ensures that the sum of all results is equal to 1.

[0132] S411: Take All the tail entities in the and the corresponding association probability p i Perform weighted summation to obtain the diagnostic history V of the abnormal phenomenon w w The first-order response vector to the candidate failure mode v :

[0133] ;

[0134] Where, For the tail entity t i Embedding;

[0135] S412: For the first-order response vector In the ripples of the anomaly w Perform iterative propagation to obtain the high-order response of the abnormal phenomenon w , the response of the abnormal phenomenon w to the candidate fault mode v is expressed as the sum of the responses of each order:

[0136] ;

[0137] Where, is the H-th order response of the abnormal phenomenon w, where H is a positive integer greater than 2;

[0138] S413: Combine the abnormal phenomenon w and the candidate failure mode v to obtain the predicted probability:

[0139] ;

[0140] Where, is the sigmoid function, is the embedding of the abnormal phenomenon w;

[0141] S414: All candidate fault modes are screened based on the predicted probabilities to obtain a recommended list of confirmed results, providing a diagnostic reference for relevant technicians. In this embodiment, a top 5 recommended list can be returned to assist relevant technicians in making diagnostic decisions.

[0142] To more intuitively demonstrate the use of the diagnosis recommendation model to perform equipment fault diagnosis recommendation tasks, we take the abnormal phenomenon of "insufficient power output" in the engine system during a certain task as an example:

[0143] 1) When the abnormal phenomenon of "insufficient power output" is used as the diagnostic input, the model will automatically link to a series of fault modes that are strongly associated with this phenomenon based on historical diagnostic information, such as Figure 3 shown.

[0144] 2) After obtaining these strongly correlated fault patterns, the model will further utilize the multi-level association relationships in the knowledge graph to conduct in-depth analysis and expansion of these patterns. Taking the iteration of the fault pattern "Injector 1 Fault" in the fault knowledge graph as an example to illustrate its two-hop propagation, as shown in the following figure: Figure 4 shown.

[0145] This process primarily propagates along two axes: One is to combine root cause analysis with the knowledge graph to identify fault modes that share a common causal relationship with the Injector 1 failure. For example, the human factor "improper maintenance procedures" and the environmental factor "corrosion in high temperature and high humidity" that cause the Injector 1 failure can also lead to "metering unit circuit breakage" and "abnormal plunger wear." Furthermore, in addition to these straightforward fault associations, the diagnostic model excels at uncovering hidden fault modes through the propagation of influence between fault modes. For example, one cause of the Injector 1 failure is wear on the injection pump camshaft. This wear in turn leads to unstable operation of the injection pump, causing injection pressure fluctuations and ultimately triggering the "Injector 2 failure." This allows for deeper fault discovery, uncovering the underlying fault from the surface to the core.

[0146] 3) After completing the multi-level association expansion of the 0-hop fault pattern, the model adds the series of fault patterns obtained during the propagation process to the candidate fault pattern list. Based on the trained model parameters, the model then calculates the predicted probability of each candidate pattern, assessing its likelihood of a diagnostic relationship under the current anomaly. Finally, the system ranks the candidate fault patterns based on these predicted probabilities and returns a list of the top five recommendations. The calculation results are shown in Table 4.

[0147] Table 4 Fault diagnosis recommendation results

[0148]

[0149] The above is a detailed introduction to the equipment fault diagnosis recommendation method based on knowledge graph and multidimensional correlation indicators provided by the present invention. In this embodiment, specific examples are used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

[0150] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in this embodiment may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown in this embodiment, but is intended to conform to the widest scope consistent with the principles and novel features disclosed in this embodiment.

Claims

1. A method for recommending equipment fault diagnosis based on knowledge graph and multi-dimensional correlation index, characterized in that: The steps include: S1: Based on the abnormal phenomena of equipment units and the fault logic of human factors, machine factors, and environmental factors, the semantic types and semantic relationships in the field of equipment failure are defined, and the equipment failure knowledge graph ontology model is constructed; S2: Obtain fault text from equipment failure cases, and convert the fault text into triple-structured data using a knowledge extraction algorithm based on the semantic types and semantic relationships in the equipment failure knowledge graph ontology model to construct an equipment failure knowledge graph; the equipment failure cases include historical fault diagnosis data; S3: Construct a multi-dimensional correlation index of abnormal phenomena corresponding to fault modes, where the fault frequency is used to record the number of times the abnormal phenomenon is diagnosed as the corresponding fault mode; Fault level is used to assess the severity of faults according to the impact of the fault mode; The impact range is used to measure the impact of the fault mode on each component in the system; the multi-dimensional correlation index is used to optimize the correlation between abnormal phenomena and fault modes, and the explicit feedback in the historical fault diagnosis data is converted into implicit feedback; the multi-dimensional correlation index is standardized to the same dimension; the standardized multi-dimensional correlation index is substituted into the Fault Mode Association Evaluation Index (FMAEI) calculation formula to measure the correlation strength between abnormal phenomena and fault modes. The calculation formula is as follows: FMAEI=α1·S f +α2·S s +α3·S i ; Where α1, α2, and α3 are the weight coefficients of the three dimensions of fault frequency, fault level, and impact range, and the sum of the weight coefficients is 1; S f , Ss, S i They are respectively the fault frequency, fault level and impact range data after normalization; S4: receiving an input abnormal phenomenon, and based on the correlation representation, obtaining a fault mode whose correlation meets a preset condition from historical fault diagnosis data as a seed fault mode; constructing a diagnosis recommendation model, starting with the seed fault mode, iteratively propagating in the fault knowledge graph to obtain a ripple set, obtaining a fault mode in the propagation process based on the ripple set, and using the seed fault mode and the fault mode in the propagation process as candidate fault modes; The association probability between the input abnormal phenomenon and the candidate fault mode is calculated according to the ripple set, each order response is accumulated according to the association probability, the predicted diagnosis probability is output, and the diagnosis result is determined.

2. The equipment fault diagnosis recommendation method based on knowledge graph and multi-dimensional correlation index according to claim 1 is characterized in that: The equipment fault knowledge graph ontology model has a multi-level structure, including: Unit layer, used to describe the spatial location and structural association information of equipment, systems and components; The logic layer is used to describe the fault characteristics and mechanisms and analyze the fault causes, including any one or more combinations of the following: human factors, machine factors, and environmental factors; Maintenance layer, used to define repair measures and solutions.

3. The equipment fault diagnosis recommendation method based on knowledge graph and multi-dimensional correlation index according to claim 1 is characterized in that: The semantic types in the equipment failure knowledge graph ontology model include any combination of the following: equipment name, system name, component name, failure mode, abnormal phenomenon, human factor, machine factor or environmental factor; the semantic relationship includes any combination of the following: composition, belonging, located, caused, induced, represented or taken.

4. The equipment fault diagnosis recommendation method based on knowledge graph and multi-dimensional correlation index according to claim 1 is characterized in that: The equipment failure case of S2 includes structured text information and unstructured text information; wherein, Structured text information includes historical fault data; Unstructured text information includes professional technical documents, troubleshooting records, and maintenance manuals.

5. The equipment fault diagnosis recommendation method based on knowledge graph and multi-dimensional correlation index according to claim 4 is characterized in that: The pre-trained BERT-BiLSTM-CRF model is used to perform named entity recognition on the unstructured text information; and the pre-trained Chinese-BERT-wwm-ext model is used to extract the relationship between fault-related entities from the unstructured text information.

6. The equipment fault diagnosis recommendation method based on knowledge graph and multi-dimensional correlation index according to claim 1 is characterized in that: In S4, a diagnosis recommendation model is constructed based on the RippleNet recommendation algorithm.

7. The equipment fault diagnosis recommendation method based on knowledge graph and multi-dimensional correlation index according to claim 4 is characterized in that: The S4 comprises the following steps: S41: calculating the correlation representation between the abnormal phenomenon and the fault mode according to the historical fault diagnosis data, and obtaining the fault mode that meets the preset conditions in the history of the input abnormal phenomenon as a seed fault mode; S42: placing the seed fault pattern into the equipment fault knowledge graph as a starting point, obtaining a related ripple set through iterative propagation, obtaining a fault pattern during the propagation process based on the ripple set, and using the seed fault pattern and the fault pattern during the propagation process as candidate fault patterns; S43: Calculating the association probability between the input abnormal phenomenon and the candidate fault mode based on the head entity and relationship of the ripple set; S44: accumulating each order response of the ripple set layer by layer according to the associated probability, accumulating each order response, combining the corresponding candidate fault mode, outputting a predicted diagnosis probability, and determining a diagnosis result according to the predicted diagnosis probability.

8. The equipment fault diagnosis recommendation method based on knowledge graph and multi-dimensional correlation index according to claim 1 is characterized in that: The step of calculating the predicted diagnosis probability between the abnormal phenomenon and the candidate failure mode in S4 includes: Obtain the 1-hop ripple set of the candidate fault mode v and the input abnormal phenomenon w in the equipment fault knowledge graph By combining the candidate failure mode v with the head entity h i and the relation r in the corresponding triple i For comparison, Each triplet (h i ,r i ,t i ) assigns an association probability p i : Where R i ∈R d×d , h i ∈R d For the relationship r i and head entity h i embedding, v is the embedding of candidate fault mode v, d is the dimension of embedding, p i Indicates v and h i The degree of correlation between The number of triplets in , softmax() is the softmax function; Pick All the tail entities in the and the corresponding association probability p i Perform weighted summation to obtain the diagnostic history V of the abnormal phenomenon w w The first-order response vector to the candidate failure mode v Where, t i ∈R d For the tail entity t i Embedding; For the first-order response vector In the ripples of the anomaly w Perform iterative propagation to obtain the high-order response of the abnormal phenomenon w The response of the abnormal phenomenon w to the candidate failure mode v is expressed as the sum of the responses of each order: Where, is the H-th order response of the abnormal phenomenon w, where H is a positive integer greater than 2; Combining the abnormal phenomenon w and the candidate failure mode v, we can get the predicted probability: Where, is the sigmoid function, w is the embedding of the abnormal phenomenon w; All candidate failure modes are screened according to the predicted probabilities to obtain a recommended list of confirmed results.

Images